File size: 11,388 Bytes
be2715b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
import numpy as np
import os 
join = os.path.join
import gc
from tqdm import tqdm
import torch
import monai, random
from segment_anything import sam_model_registry
from dataloader.sam_transforms import ResizeLongestSide
from dataloader.dataloader import sam_dataloader
from utils.SurfaceDice import multiclass_iou


def fit(cfg,
        sam_model,
        train_loader,
        valid_dataset,
        optimizer,
        criterion,
        model_save_path):
    """
    Function to fit model
    """
    
    best_valid_iou3d = 0
    
    device = cfg.base.gpu_id
    num_epochs = cfg.train.num_epochs
    
    for epoch in range(num_epochs):
        sam_model.train()
        
        epoch_loss = 0
        valid_iou3d = 0
        
        print(f"Epoch #{epoch+1}/{num_epochs}")
        for step, batch in enumerate(tqdm(train_loader, desc='Model training', unit='batch', leave=True)):

            """
            Load precomputed image embeddings to ease training process
            We also load mask labels and bounding boxes directly computed from ground truth masks
            """
            image, true_mask, boxes = batch['image'], batch['mask'], batch['bboxes']
            sam_model = sam_model.to(f"cuda:{device}")
            image = image.to(f"cuda:{device}")
            true_mask = true_mask.to(f"cuda:{device}")

            """
            We freeze image encoder & prompt encoder, only finetune mask decoder
            """
            with torch.no_grad():
                """
                Compute image embeddings from a batch of images with SAM's frozen encoder
                """
                encoder = torch.nn.DataParallel(sam_model.image_encoder, device_ids=[3, 2, 1, 0], output_device=device)
                encoder = encoder.to(f"cuda:{encoder.device_ids[0]}")
                sam_model = sam_model.to(f"cuda:{encoder.device_ids[0]}")
                image = image.to(f"cuda:{encoder.device_ids[0]}")
                image = sam_model.preprocess(image[:, :, :])
                image_embedding = encoder(image)

                """
                Get bounding boxes to make segmentation prediction
                We follow the work by Jun Ma & Bo Wang in Segment Anything in Medical Images (2023)
                to get bounding boxes from the masks as the boxes prompt for SAM
                """
                box_np = boxes.numpy()
                sam_trans = ResizeLongestSide(sam_model.image_encoder.img_size)
                box = sam_trans.apply_boxes(box_np, (true_mask.shape[-2], true_mask.shape[-1]))
                box_torch = torch.as_tensor(box, dtype=torch.float, device=f"cuda:{device}")
                if len(box_torch.shape) == 2:
                    box_torch = box_torch[:, None, :] # (B, 1, 4)
                
                """
                Encode box prompts information with SAM's frozen prompt encoder
                """
                prompt_encoder = torch.nn.DataParallel(sam_model.prompt_encoder, device_ids=[0,1,2,3], output_device=device)
                prompt_encoder = prompt_encoder.to(f"cuda:{prompt_encoder.device_ids[0]}")
                box_torch = box_torch.to(f"cuda:{prompt_encoder.device_ids[0]}")
                sparse_embeddings, dense_embeddings = prompt_encoder(
                    points=None,
                    boxes=box_torch,
                    masks=None,
                )

            """
            We now finetune mask decoder
            """
            sam_model = sam_model.to(f"cuda:{device}")
            predicted_mask, iou_predictions = sam_model.mask_decoder(
                image_embeddings=image_embedding.to(f"cuda:{device}"), # (B, 256, 64, 64)
                image_pe=sam_model.prompt_encoder.get_dense_pe(), # (1, 256, 64, 64)
                sparse_prompt_embeddings=sparse_embeddings, # (B, 2, 256)
                dense_prompt_embeddings=dense_embeddings, # (B, 256, 64, 64)
                multimask_output=False,
            ) # -> (B, 1, 256, 256)

            predicted_mask = predicted_mask.to(f"cuda:{device}")
            true_mask = true_mask.to(f"cuda:{device}")
            loss = criterion(predicted_mask, true_mask)
            
            """
            Upgrade model's params
            """
            optimizer.zero_grad(set_to_none=True)
            loss.backward()
        
            clip_value = 1 # Clip gradient
            torch.nn.utils.clip_grad_norm_(sam_model.mask_decoder.parameters(), clip_value)
        
            optimizer.step()
            epoch_loss += loss.item()
        
        """
        Validation step with IoU as the metric
        """
        with torch.no_grad():
            valid_iou3d = eval_iou(sam_model,
                                 valid_dataset,
                                 device=device)
    
        epoch_loss /= ((step + 1) * len(train_loader))
        print(f'Loss: {epoch_loss}\n---')
        
        """
        Save best model
        """
        if best_valid_iou3d < valid_iou3d:
            best_valid_iou3d = valid_iou3d
            torch.save(sam_model.state_dict(), join(model_save_path, f'{cfg.base.best_valid_model_checkpoint}{cfg.base.random_seed}.pth'))
        
        print(f"Valid 3D IoU: {valid_iou3d*100}")
        print('=======================================')
            
    print(f"Best valid 3D IoU: {best_valid_iou3d*100}")

def eval_iou(sam_model,
            loader,
            device):
    """
    We use IoU to evalute 3D samples.
    
    For 3D evaluation, we first concatenate 2D slices into 1 unified 3D volume and pass into model
    However, due to limited computational resources, we could not perform 3D evaluation in GPU.
    Hence, I set up to perform this function completely on CPU. 
    If you have enough resources, you could evaluate on multi-gpu the same as in training function.
    """
    gc.collect()
    torch.cuda.empty_cache()
    torch.cuda.reset_peak_memory_stats()
    
    iou_score = 0
    num_volume = 0
    for _, batch in enumerate(tqdm(loader.get_3d_iter(), leave=False)):
        """
        Load precomputed embeddings, mask labels and bounding boxes computed directly from ground truth masks
        """
        image, true_mask, boxes = batch['image'], batch['mask'], batch['bboxes']
        image = image.to(f"cpu")
        true_mask = true_mask.to(f"cpu", dtype=torch.float32)

        """
        Compute image embeddings
        """
        sam_model = sam_model.to(f"cpu")
        image = image.to(f"cpu")
        image = sam_model.preprocess(image[:, :, :])
        image_embedding = sam_model.image_encoder(image)

        """
        Get bboxes
        """
        box_np = boxes.numpy()
        sam_trans = ResizeLongestSide(sam_model.image_encoder.img_size)
        box = sam_trans.apply_boxes(box_np, (image_embedding.shape[0], image_embedding.shape[1]))
        box_torch = torch.as_tensor(box, dtype=torch.float32, device=device)
        if len(box_torch.shape) == 2:
            box_torch = box_torch[:, None, :] # (B, 1, 4)
        
        """
        Prompt encoder component
        """
        box_torch = box_torch.to(f"cpu")
        sparse_embeddings, dense_embeddings = sam_model.prompt_encoder(
            points=None,
            boxes=box_torch,
            masks=None,
        )

        """
        Mask decoder component
        """
        sam_model = sam_model.to(f"cpu")
        mask_segmentation, iou_predictions = sam_model.mask_decoder(
            image_embeddings=image_embedding.to(f"cpu"), # (B, 256, 64, 64)
            image_pe=sam_model.prompt_encoder.get_dense_pe(), # (1, 256, 64, 64)
            sparse_prompt_embeddings=sparse_embeddings, # (B, 2, 256)
            dense_prompt_embeddings=dense_embeddings, # (B, 256, 64, 64)
            multimask_output=False,
        ) # -> (B, 256, 256)
        
        """
        Transform prediction and evaluate
        """
        true_mask = true_mask.to("cpu")
        medsam_seg_prob = torch.sigmoid(mask_segmentation)
        medsam_seg = (medsam_seg_prob > 0.5).to(dtype=torch.float32)
        iou_score += multiclass_iou((true_mask>0).to(dtype=torch.float32), (medsam_seg>0).to(dtype=torch.float32))
        num_volume += 1
    return iou_score.cpu().numpy()/num_volume
    
def medsam_3d(yml_args, cfg):
    """
    Training warm up
    """
    torch.multiprocessing.set_start_method('spawn')
    
    random.seed(cfg.base.random_seed)
    np.random.seed(cfg.base.random_seed)
    torch.manual_seed(cfg.base.random_seed)
    torch.cuda.manual_seed(cfg.base.random_seed)

    torch.backends.cudnn.deterministic = True

    gc.collect()
    torch.cuda.empty_cache()
    torch.cuda.reset_peak_memory_stats()

    """
    General configuration
    """
    img_shape = (3, 1024) # hard settings image shape as 3 x 1024 x 1024
    model_save_path = join("./work_dir", 'SAM-ViT-B')
    os.makedirs(model_save_path, exist_ok=True)

    print(f"Fine-tuned SAM (3D IoU) in {cfg.base.dataset_name} with {cfg.train.optimizer}, LR = {cfg.train.learning_rate}")

    """
    Load SAM with its original checkpoint 
    """
    sam_model = sam_model_registry["vit_b"](checkpoint=cfg.base.original_checkpoint)

    """
    Load precomputed embeddings
    """    
    train_loader, _, _, valid_dataset, test_dataset = sam_dataloader(cfg)

    """
    Optimizer & learning rate scheduler config
    """
    if cfg.train.optimizer == 'sgd':
        optimizer = torch.optim.SGD(sam_model.mask_decoder.parameters(),
                                    lr=float(cfg.train.learning_rate),
                                    momentum=0.9)
    elif cfg.train.optimizer == 'adam':
        optimizer = torch.optim.Adam(sam_model.mask_decoder.parameters(), 
                                     lr=float(cfg.train.learning_rate),
                                     weight_decay=0,
                                     amsgrad=True)
    elif cfg.train.optimizer == 'adamw':
        optimizer = torch.optim.AdamW(sam_model.mask_decoder.parameters(), 
                                     lr=float(cfg.train.learning_rate),
                                     weight_decay=0)
    else:
        raise NotImplementedError(f"Optimizer {cfg.train.optimizer} is not set up yet")
    
    """
    Loss function
    In this work, we use a combination of Dice and Cross Entropy Loss to measure SAM's loss values.
    """
    criterion = monai.losses.DiceCELoss(sigmoid=True, 
                                        squared_pred=True, 
                                        reduction='mean')

    """
    Train model
    """
    if not yml_args.use_test_mode:
        fit(cfg,
            sam_model=sam_model,
            train_loader=train_loader,
            valid_dataset=valid_dataset,
            optimizer=optimizer,
            criterion=criterion,
            model_save_path=model_save_path)

    """
    Test model
    """
    with torch.no_grad():
        sam_model_test_iou = sam_model_registry["vit_b"](checkpoint=join(model_save_path, f'{cfg.base.best_valid_model_checkpoint}{cfg.base.random_seed}.pth'))
        sam_model_test_iou.eval()
        test_iou_score = eval_iou(sam_model_test_iou,
                                    test_dataset,
                                    device=cfg.base.gpu_id)
        print(f"Test 3D IoU score after training with {cfg.train.optimizer}(lr = {cfg.train.learning_rate}): {test_iou_score *100}")